Thoughts

Terminal Use may be a lower-dimensional projection of Computer Use

What actually makes an evaluation valuable?

For practical computer use, perhaps a11y should be the first choice and VLM the fallback.

Today's end-to-end Computer Use evaluations are too heavy and expensive; we need a lower-level capability evaluation.

Many software problems only appear through real use, and agent harnesses make that pattern obvious.

Embodied AI and CUA are the same thing in two different worlds — and the digital world model should be easier.

CUA training infrastructure is three lines: Data, Eval, Train — connected by Gym.

The missing product in the AI agent ecosystem isn't another framework — it's a client app designed for users.

In the age of AI-driven output explosion, promotion is no longer optional — it's a core skill.

Agent swarm communication is unsolved at the product level — markdown files and polling are still the default.

The bottleneck of long-horizon agentic tasks isn't prompting — it's that chat-trained models aren't built to act autonomously.

CUA's capabilities are essentially three layers: thinking, discrete operations, and continuous operations.

Embodied AI and CUA share the same core challenges — world modeling, rule discovery, and experience memory.

CLIzing everything — as agents grow stronger, more apps and websites should open interfaces for them.

The fundamental difference between agentic and chat paradigms is the ability to explore through tool use.